[sixbrx]

First, correctness first. Second, correlated subqueries in selects really aren't noticably slow at all in my experience (Postgres, Oracle) when they're really needed, meaning alternatives are as slow or slower. Especially compared to the alternative of actually doing a cartesian product across independent children of any size. I don't think the optimizer could help much with avoiding that cartesian product to be actually realized, either - the grouping after that join is going to be a big sort in general with results that depend on the distribution of child values on a row-by-row basis. But in any case it would give the wrong answers for both children regardless of speed and memory used!